Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Albert]

This thread is sort of an offshoot of the thread about digital attenuation. After reading about the issues regarding simple attentuation of a signal in the digital domain, I'm curious about what must be an even more involved process, that of mixing in the digital domain.

With mixing, not only do you have attenuation and gain, but you have things like fx plugins, sometimes sending and returning to external analog or digital processing boxes, and finally, very importantly, digital summing of all signals that have been through the previous processes.

The math involved in doing attenuation properly is staggering, based on what I read in the other thread. So what is really happening to a group of digital tracks in a DAW that must go through attentuation and gain, plugins, and summing? Is there anything left of the original track, meaning the original ones and zero's? I'm particularily interested in how the numbers work when summing digital tracks.

I apologize if this topic has already been covered here. Thank you in advance.

[danlavry]

Albert wrote on Thu, 23 December 2004 15:38

This thread is sort of an offshoot of the thread about digital attenuation. After reading about the issues regarding simple attentuation of a signal in the digital domain, I'm curious about what must be an even more involved process, that of mixing in the digital domain. With mixing, not only do you have attenuation and gain, but you have things like fx plugins, sometimes sending and returning to external analog or digital processing boxes, and finally, very importantly, digital summing of all signals that have been through the previous processes. The math involved in doing attenuation properly is staggering, based on what I read in the other thread. So what is really happening to a group of digital tracks in a DAW that must go through attentuation and gain, plugins, and summing? Is there anything left of the original track, meaning the original ones and zero's? I'm particularily interested in how the numbers work when summing digital tracks. I apologize if this topic has already been covered here. Thank you in advance.

This thread is sort of an offshoot of the thread about digital attenuation. After reading about the issues regarding simple attentuation of a signal in the digital domain, I'm curious about what must be an even more involved process, that of mixing in the digital domain.

With mixing, not only do you have attenuation and gain, but you have things like fx plugins, sometimes sending and returning to external analog or digital processing boxes, and finally, very importantly, digital summing of all signals that have been through the previous processes.

The math involved in doing attenuation properly is staggering, based on what I read in the other thread. So what is really happening to a group of digital tracks in a DAW that must go through attentuation and gain, plugins, and summing? Is there anything left of the original track, meaning the original ones and zero's? I'm particularily interested in how the numbers work when summing digital tracks.

I apologize if this topic has already been covered here. Thank you in advance.

This is a big question. I’ll try to make it reasonably short:

The math for gain and attenuation is not all that difficult. If you know where the noise floor is for a given track, than if you increase the gain by Xdb, the noise will rise by X dB. For attenuation it is a bit more difficult but not too bad…

In all cases, from a purist standpoint, the best you can do is to record each track with the music peaks as close to full scale as possible, but never to the point of clipping. That way, you utilize digital audio in the best possible way. That way you do not need to amplify signals (and with it amplify their noise floor). If you need to attenuate, you will end up about where you would be recording further down from the full scale, no real loss. A weak signal in digital audio is simply closer to the noise floor. Often you such signals have somewhat “masked” by a stronger signal within the mix.

Mixing digital is not so bad. Say you add 2 identical 1KHz tones (IN PHASE), each at say -1dBFS (dB from full scale) Say each has a -90dBFS noise. The added outcome is doubling of the signal. Double means +6dB so you now have +5dB signal. What happened to the noise? Unlike the signals – an addition of 2 identical waves, the noise of each channel is different. Say it is random, than the total noise will go up by only 3dB, to -87dBFS.  Of course a +5dBFS is over full scale. Say we attenuate it by 6dB. The signal is now again at -1dBFS, the noise went down to -93dBFS. This is an improvement.

This is not a typical case. How often do you add 2 identical signals? It is safer and more realistic to assume that  2 signals (at the same overall levels) are very un related to each (low correlation) and than the addition will not double, but increase the overall level by some amount, such as 3dB, just like the noise. So an addition and rescaling will not alter the signal to noise ratio. It is better than “attenuation only” because you not just attenuating, you are also adding a signal.

It does get a bit more complicated when the signals are to be mixed at unequal relative strength, but not impossible to comprehend. The best way to be safe is to have a reasonably low noise on the original track and a wide enough mix bus (bits).

Analog attenuation is cleaner than digital one. Analog gain is cleaner than a digital one. But analog mix bus (adding of signals) also has its own peculiar set of problems…

Regards
Dan Lavry

[Albert]

Okay, so far so good. But you have all these streams of 1's and 0's, each having started with a series of 1's and 0's in the original recording. They are then altered when attenuated or gain is applied, altered again by menas of plugins, and finally, these streams of numbers must be combined.

How does that combination take place? Let's say you have 24 tracks of digital audio, each track it's own distinct data stream. This is then reduced down to two data streams, the stereo mix. How is all the 24 channels of data combined into two? Is there numerical averaging going on, or are samples taken at the current bit rate?

I'm just curious how this mathematical process takes place, and why I don't like the sound of it as compared to an analog mix, or rather analog summing.

[Ethan Winer]

Albert,

> How does that combination take place? Let's say you have 24 tracks of digital audio <

It's really simple.

Seriously, mixing two tracks (or any number of tracks) is simple addition. At any given moment - the current sample for all tracks - the DAW software simply adds the numbers for all tracks together. The resulting number defines the loudspeaker's displacement, either moving outward (positive numbers) or inward (negative). Gain changes are also very simple, only in that case each track's sample number is multiplied by a constant amount instead of being added to other track sample numbers.

I also feel compelled to point out that the artifacts some people complain about during digital summing and gain changing are way overstated. What it really comes down to is distortion, and the added distortion with each math operation is really REALLY tiny in the grand scheme of things. Certainly many orders of magnitude below the distortion of any microphone or loudspeaker. Even after many such operations in a row.

--Ethan

[Albert]

Ethan, thanks for your response.

If it is simple addition, then why do the various DAW's sound different? Seems like they should all sound exactly the same.

Is there a book I can read on this subject?

[Barry Hufker]

A wonderful book on this subject is "Principles of Digital Audio" by Ken Pohlman.

Barry

[sdevino]

danlavry wrote on Thu, 23 December 2004 19:59

In all cases, from a purist standpoint, the best you can do is to record each track with the music peaks as close to full scale as possible, but never to the point of clipping. That way, you utilize digital audio in the best possible way. That way you do not need to amplify signals (and with it amplify their noise floor). If you need to attenuate, you will end up about where you would be recording further down from the full scale, no real loss. A weak signal in digital audio is simply closer to the noise floor. Often you such signals have somewhat ?masked? by a stronger signal within the mix.

Of course when doing this your analog front end is probably running with ver y high levels of gain pushing the analog clip limits and creating lots of distortion and noise. So this setup is purely a hypothetical best case for digital.

Quote:

Mixing digital is not so bad. Say you add 2 identical 1KHz tones (IN PHASE), each at say -1dBFS (dB from full scale) Say each has a -90dBFS noise. The added outcome is doubling of the signal. Double means +6dB so you now have +5dB signal. What happened to the noise? Unlike the signals ? an addition of 2 identical waves, the noise of each channel is different. Say it is random, than the total noise will go up by only 3dB, to -87dBFS.  Of course a +5dBFS is over full scale. Say we attenuate it by 6dB. The signal is now again at -1dBFS, the noise went down to -93dBFS. This is an improvement.

... and these idenetical issues exist in all analog mixing as well.

Quote:

Analog attenuation is cleaner than digital one. Analog gain is cleaner than a digital one. But analog mix bus (adding of signals) also has its own peculiar set of problems?

What do you mean by "cleaner" in this case Dan? Can you clarify?

My goal in commenting here was to clarify, not refute.

Steve

[RKrizman]

Albert wrote on Fri, 24 December 2004 15:44

If it is simple addition, then why do the various DAW's sound different? Seems like they should all sound exactly the same.

What makes you think they all sound different?  AFAIK,  they all sound the same if you just compare their summing and level change functions.

I mean, why wouldn't they?

-R

[Ethan Winer]

Albert,

> If it is simple addition, then why do the various DAW's sound different? <

I agree with RK. All DAWs do sound the same. If they don't then it's a coding error, or a sound card driver issue, or something else like that.

--Ethan

[Bob Olhsson]

RKrizman wrote on Sun, 26 December 2004 01:04

... What makes you think they all sound different?  AFAIK,  they all sound the same if you just compare their summing and level change functions. I mean, why wouldn't they? -R

They would all need to use the exact same math, precision, gain structure and dither!

Doing gain changes in a first class fashion reduces the number of possible features. Because most people (including too many misinformed developers) assume digital gain changes to be trivial, developers are inclined to spend their processing power elsewhere. Roger Lagadec presented a fascinating paper to the AES about the difficulty of creating a truly transparent digital volume control.

[Timeline]

This topic is one that I feel could be very useful to many of us. Here are some of my
questions I have wondered about for a long time.

1. Would mixing multiple track transient sources not be a reason to seek out systems with higher sample and or bit rates?   In-other-words, do complex waveforms actually require higher bandwidths to be properly reproduced compared to an individual track?  

2. If you record your tracks near the last bit possible, how might a digital mix buss then be an improved landscape for combining complex waveforms, i.e. mixing, of drums, percussion and all musical sources at once over analog? Compare digital combining to simply sending out instruments on an individual channels and mixing in analog where possible 'good sounding musical harmonics' are generated?

3. When your individually recorded multi-channel digital audio is reproducing itself at unity gain at the IO's output, are these levels in the analog devices of the IOs causing more problems in general to clarity due to slewing or lack of headroom? For instance MOTU has no facility to reduce it's OP levels in their IO's other than digitally reducing the level in the DAW. How may IO manufacturers have audio trims to solve this issue and have manufacturers really taken into consideration these higher OP levels for clean reproduction of sound?

Thanks

Gary Brandt
Producer/Engineer

[Bob Olhsson]

Every single one of these things is entirely a question of the implementation of specific equipment and software. While most people in music would love a simple formula for achieving high quality sound, the unpleasant reality is that there isn't any.

The closest thing I've found to a simple formula is investing in the very highest quality monitoring and D to A conversion you can put together. Only then does it become obvious how irrelevant most popular generalizations are to achieving truly great sounding audio.

[sdevino]

Timeline wrote on Sun, 26 December 2004 13:10

1. Would mixing multiple track transient sources not be a reason to seek out systems with higher sample and or bit rates?   In-other-words, do complex waveforms actually require higher bandwidths to be properly reproduced compared to an individual track?

The complexity of the transients does not effect required sample rate. ALl the transient information that makes up a 20 kHz bandwidth signal are within that 20kHz band. Transients that occur "between the sample points" represent frequency content that is higher than accepted audio frequencies.

So you have to settle the argument regarding audio bandwidth requirements before you can answer this question completely.

Steve

[ammitsboel]

Bob Olhsson wrote on Sun, 26 December 2004 18:32

The closest thing I've found to a simple formula is investing in the very highest quality monitoring and D to A conversion you can put together. Only then does it become obvious how irrelevant most popular generalizations are to achieving truly great sounding audio.

That was nicely put!

Best Regards,

[bobkatz]

Timeline wrote on Sun, 26 December 2004 13:10

1. Would mixing multiple track transient sources not be a reason to seek out systems with higher sample and or bit rates?   In-other-words, do complex waveforms actually require higher bandwidths to be properly reproduced compared to an individual track?

If all you are doing is mixing, it's not possible to create INTERSAMPLE peaks (additional peaks between the samples, or "higher bandwidth"). The material was prefiltered to the Nyquist rate when it was recorded, and you'll create nothing new there.

When the mixed material, which may include some equalization, for example, feeds the outside world, like a DAC, it's a good idea to use an oversampled peak meter, or, drop your levels to -3 dBFS measured on a simple, ordinary peak meter. Within the DAC, or any SRC, for that matter, the oversampling process, and filtering inside the DAC, can produce intersample peaks that are higher than what a simple digital meter reveals. But note: You do not have to worry about this "creation of extra bandwidth" UNTIL you SRC or until you feed a DAC.

What about processsing? Filters in your chain can produce higher output levels than their inputs (but not intersample peaks, so no worries there as long as you remain in the digital domain and your digital meter does not show an over).

However, non-linear processors like compressors can create serious distortion if left within "single" sample rate. So, if you're going to be compressing digitally, it's a good idea to think about using "double" sample rate to begin with! Some compressors internally double sample and then internally sample back down. If this is handled well (and it's a BIG CPU HOG!) then conceivably you can process and mix at the single sample rate (44 or 48); your mileage may vary depending on the quality of the digital compressor.

Quote:

2. If you record your tracks near the last bit possible, how might a digital mix buss then be an improved landscape for combining complex waveforms, i.e. mixing, of drums, percussion and all musical sources at once over analog? Compare digital combining to simply sending out instruments on an individual channels and mixing in analog where possible 'good sounding musical harmonics' are generated?

It's a tradeoff. Just make sure you're not fooling yourself because of level differences or anything else. Make tests at matched gain and decide for yourself. You're entering the world of noise, buzz, hum, hiss, distortion, and transparency loss. As well as (conceivably) perceived gain due to the quality of the distortion that is added, or the "softening" of the sound due to the loss of transparency. It's up to you to decide if the TOTAL "cure" sounds better than the "TOTAL" disease. Not every analog console you can dream up is a sonic winner in this regard, and it really pays to use superior converters with superior clocking as well, if you are going to be mixing analog.

I have a client who mixes through an SSL G or E, and he doesn't fool himself that there isn't a loss. In fact, for jazz music, there isn't much desirable "character" in the SSL that he or I can hear. He also bypasses the entire center section, which helps a lot. However, he loves the SSL automation, and as I said above, there are lots of advantages of using good ol' fashioned analog compressors once you've absorbed the hit of that first D/A conversion. (Perhaps this is a justification for compressing  in the analog domain while tracking, but that's another subject---I don't think you can hit on the nose how much compression you need while you are tracking not in the context of mixing)

Quote:

3. When your individually recorded multi-channel digital audio is reproducing itself at unity gain at the IO's output, are these levels in the analog devices of the IOs causing more problems in general to clarity due to slewing or lack of headroom? For instance MOTU has no facility to reduce it's OP levels in their IO's other than digitally reducing the level in the DAW. How may IO manufacturers have audio trims to solve this issue and have manufacturers really taken into consideration these higher OP levels for clean reproduction of sound?

I think you are hitting a few nails on the head here. Internal headroom in converters is often not well taken care of. That's why I recommended -3 dBFS peak maximum when converting, unless you have a well-designed converter.

HTH,

BK

[Timeline]

Thanks Bob,

"I think you are hitting a few nails on the head here. Internal headroom in converters is often not well taken care of. That's why I recommended -3 dBFS peak maximum when converting, unless you have a well-designed converter."

Do you mean -3 on my IO's peak meter?

"If all you are doing is mixing, it's not possible to create INTERSAMPLE peaks (additional peaks between the samples, or "higher bandwidth"). The material was prefiltered to the Nyquist rate when it was recorded, and you'll create nothing new there. "

So your saying even if you record at 96K or above nyquist filter would strip harmonics and no interaction above 20K even though the audio bandwidth is higher could exist?  That doesnt make sense if you have audio there. It can't be all noise.

Lavry said  96K SR was well beyond the filter points and 60K SR was clear.

I don't get it.  

[bobkatz]

Timeline wrote on Sun, 26 December 2004 20:41

Thanks Bob, "I think you are hitting a few nails on the head here. Internal headroom in converters is often not well taken care of. That's why I recommended -3 dBFS peak maximum when converting, unless you have a well-designed converter." Do you mean -3 on my IO's peak meter?

Yes, of course that's extremely conservative. If you have a sharp eye, -1 dBFS peak will also do. But you have tons of "footroom" at 24 bit, why not just be safe and be done with it.

Quote:

BK said: If all you are doing is mixing, it's not possible to create INTERSAMPLE peaks (additional peaks between the samples, or "higher bandwidth"). The material was prefiltered to the Nyquist rate when it was recorded, and you'll create nothing new there. and you replied: So your saying even if you record at 96K or above nyquist filter would strip harmonics and no interaction above 20K even though the audio bandwidth is higher could exist?  That doesnt make sense if you have audio there. It can't be all noise.

No, that's not what I said. I said that digital mixing would not create any additional bandwidth above Nyquist. Since the 96K sample rate can conceivably contain harmonics of the music up to its Nyquist frequency of 48 kHz, then if they are there in the recording, the digital mixing will preserve them, not adding or taking away from them.

You should see what kinds of non-musical garbage above 20 kHz my Cedar Retouch color graph display shows in a number of "96 kHz" recordings. Remember, all the buzz and harmonics of tics and pops and electrical noises end up there along with whatever harmonics of the music might be up there. Oh well, it's inaudible, anyway?  Who cares about a few burnt out tweeters?  (My tweeters have never burned out so it's an academic question anyway).

BK

[Timeline]

OK Bob,

Thanks very much for your fine advise on IO peaks and I will do exactly that on my next sessions.

After reading Dans threads and your comments, I thought it would be a good idea to freshen my memories of this test I did once on a ProTools rig.

So, I just now finished convertomg a tune of 54 tracks of 96k audio to 48K that I had been working on today. it was mostly live instruments. Rock stuff.  

The G5 2.5ghz computer headroom was fine at 96k and I thought it would be a good one tune to use because of the many vocal tracks and rock percussive rhythms.

I used the internal clock of one of my MOTU HD192's as sync master per reading here that this would likely be best for the conversion.  I used the DAW's best algorithm available as well which doubled the conversion time.

When it finished,  about an hour later, I A/B'd the sound of the tracks mixed with identical eq's and plugs in exactly the same mannor I had been listening all day.

My setup was that I fed multitrack analog from the DA's IO's through the same API mixer at calibrated levels to an L2 and then to monitor.  I listened on headphones, MA1 Dynaudio monitors and M-audio BX8 for a lo-fi comparison.

What I heard was: less top on the rhythm & acoustic guitars.  Silkiness between symbols snare and hard transient percussion like tamb and shakers lost snap as well as most of the other instruments. Vocals became a bit more grainy.
Overall just less hi-fi on the top end and the mix sounded clean but with less separation in the mids. Similar to taking 15k and rolling off 2 db shelf.

Also the low end didn't seem quite as nice although I don't understand why.

Here's the thing.  At all times the entire all day session and this 96K to 48K comparison was done monitoring THROUGH an L2 set to 44.1  16bit.

Since Nyquist rolls everything off in the L2 from either playbacks above 20khz,  why did it still sound cleaner with 96K on the original audio tracks out analog?

Consider also that I did this same test at O'Henry Studios Burbank about 2 years ago on a ProTools HD system and it rendered the same results to my ear and the tech in the room.

Would you agree for now that with todays commonly used collection of digital audio gear it might be safer to continue 96K recordings to preserve our projects for future remixes for a time when better gear might be available considering we don't really know 'what's inside the box' and why these sound changes appear in the most popular equipment used in todays studios?   This is still gray area to my ear.

Thanks

Gary Brandt

[Albert]

Gary, my response is based on personal experience, not scientific knowledge:

In my opinion, if you had recorded the original tracks at 48k it would sound better than recording at 96k and then SRC'ing down to 48k. Likewise, if you had played the mix back at 96k, gone analog through good converters and re-recorded it through good converters on another machine running at 48k it would sound better than using SRC to reduce the sample rate. Also, I think to be entirely objective, your test would require that the audio be played back directly to analog from 96k and then directly to analog at 48k. Putting the L2 at 44/16 in there would seem to me to skew the listening test because you are adding the variable of how it handles the calculations to 16/44 from two different sample rates.

Most likely, I have not had access to the very finest in SRC software/gear, but I've never been entirely happy with PT, DP, Peak, BarbaBatch, or even my hardware L2, although that has been pretty good.

But SRC is probably a subject for a whole 'nother thread.

Getting back to mixing in the box, I'd like to revisit the comment about summing being just math, and then Bob O's reply. If indeed summing is simple addition, then why would this be implemented differently in different DAW's? Addition is addition.

And I do hear differences between DAW's. Often very very subtle, but there nevertheless. Even playing back two track stereo mixes between different software can produce slightly differnet tone or levels.

[Bob Olhsson]

While addition IS addition, there are limits to how much you can add before running out of headroom. It's a SYSTEM. This must ALWAYS be borne in mind!

How the application systematically handles the level changes required can be all over the map in terms of the number and accuracy of calculations and how the level changes are dithered back to something that can be written to a file or sent to an output.

[bobkatz]

Timeline wrote on Tue, 28 December 2004 01:28

OK Bob, Thanks very much for your fine advise on IO peaks and I will do exactly that on my next sessions.

Maaaa Pleazur

Quote:

So, I just now finished convertomg a tune of 54 tracks of 96k audio to 48K that I had been working on today. it was mostly live instruments. Rock stuff.   SNIP I used the internal clock of one of my MOTU HD192's as sync master per reading here that this would likely be best for the conversion.

Well, if you are doing a digital sample rate conversion from file to new file, the MOTU operates without any clock or real time audio at all. You can do this without a clock. Unless Motu complains because its audio system wants a clock, which is just a subsystem, not relevant to the other subsystem doing the SRC that doesn't even need an audio interface.

Quote:

I used the DAW's best algorithm available as well which doubled the conversion time.

Were you able to produce 32 bit float output files?  If not, did you tell it to dither down from 32 to 24 after doing the conversion from 96 khz to 48 kHz?

Quote:

What I heard was: less top on the rhythm & acoustic guitars.  Silkiness between symbols snare and hard transient percussion like tamb and shakers lost snap as well as most of the other instruments. Vocals became a bit more grainy. Overall just less hi-fi on the top end and the mix sounded clean but with less separation in the mids. Similar to taking 15k and rolling off 2 db shelf.

There is always some loss. But it's not due to the sample rate. It's due to several variables:

1) Consider whether your D/A converter's upsampling ratios and internal filtering are giving your less than optimum picture or presentation at the two different sample rates

2) Consider whether MOTU's algorithm was as well implemented as it could be (see dither notes above) and/or is as good as a $4000 Weiss or DCS or Lavry or other sample rate converter

3) And even with all the above considered, since every DSP in line is cumulative, it is possible, that even with the VERY best of the above, that that single generation of additional filtering to reduce the rate, on top of the original recording, was enough to take it down hill. However, this has not been my experience to my ears in straight SRCs done with extremely high quality algorithms and then reproduced by a D/A converter whose upsampling path at 48 khz is as good-sounding as at 96K.

And a question: Is there a technical reason why you reduced the rate to 48K now rather than mix at 96K, and retain that resolution until the last minute when and ONLY in the mastering it gets reduced to 44? That reduces one more huge filtering and calculatoin step.

Quote:

Here's the thing.  At all times the entire all day session and this 96K to 48K comparison was done monitoring THROUGH an L2 set to 44.1  16bit.

I might be a little obtuse, but I don't understand the block diagram. How did you feed the L2 a 96K or 48K source and get it to say "4416" at all? You were using the D/A converter in the L2?

Quote:

Since Nyquist rolls everything off in the L2 from either playbacks above 20khz,  why did it still sound cleaner with 96K on the original audio tracks out analog?

Can we leave that question open until we sort out the other data?  For many years I heard a BIG difference between 96 and 48 (especially 44) until I updated to converters whose philosophy is to more properly upsample the lower rates. Also, for many years I heard a BIG difference between 96 sources and downsampled results until I got the Weiss SRC. One pass through the Weiss, monitored through a good DAC, is the least degraded SRC I've heard so far. I'm sure the DCS and a few others pass that test as well.

However, two passes through the Weiss and you start to hear it. It's important to separate the issues, though. Take a 96 down to 44 and then back up to 96 and it sounds very very close to the original. This helps to eliminate the variables of whether the sample rate itself is poorly reproduced on a DAC. Calculations in the digital world have gotten better than I think MOTU at least can handle. Have you tried Barbabatch? You might be in for a surprise. NO, I have not personally done a shootout yet between Barbabatch and MOTU, but I have a lot of confidence in BB's file integrity. BB also allows you to either make 32 bit float files (which unfortunately MOTU cannot handle to my knowledge) or dither down to 24 within the SRC algorithm.

Quote:

Consider also that I did this same test at O'Henry Studios Burbank about 2 years ago on a ProTools HD system and it rendered the same results to my ear and the tech in the room.

Not surprised. One of the reasons we are arguing for the 96 K sample rate it is a lot easier (less costly, less lossy) to deal with in the face of the mediocre DSP that we encounter. That's why I suggested you mix and stay at 96K for as long as possible and then let the mastering engineer do the one and only conversion to 44 and to 16 at the very very end of the game. You will get better results, I doubt anyone would argue with you.

It is a separate issue from whether 96 is "better" than 48. In a perfectly designed world, they should be virtually indistinguishable. But perfect DACs and perfect filters hardly exist.

Quote:

Would you agree for now that with todays commonly used collection of digital audio gear it might be safer to continue 96K recordings to preserve our projects for future remixes for a time when better gear might be available considering we don't really know 'what's inside the box' and why these sound changes appear in the most popular equipment used in todays studios?   This is still gray area to my ear.

I'm 100% in favor of it for that reason and also because of the reasons of cumulative degradation. It just takes longer to reach the bottom of the hill if you start higher on the hill  

After a few mixes and compressors, and equalizers and filters are thrown in, the inevitable degradation begins.

Hope this helps, too,



Bob

[bobkatz]

Albert wrote on Tue, 28 December 2004 12:18

Gary, my response is based on personal experience, not scientific knowledge: In my opinion, if you had recorded the original tracks at 48k it would sound better than recording at 96k and then SRC'ing down to 48k.

If your goal is to mix at 48K, I would tend to agree, the losses of the SRC conversion completely cancel out any advantages of an original recording at 96K. However, why not mix at 96K in the first place?

Quote:

Likewise, if you had played the mix back at 96k, gone analog through good converters and re-recorded it through good converters on another machine running at 48k it would sound better than using SRC to reduce the sample rate.

Now I think this is the reaction of someone who has not experienced the most transparent of SRC conversions. It's been a good 5-10 years since D/A/D conversion sounds better, transparent, more open, and purer than a given SRC. Ask 10 mastering engineers. Ask Bob Ludwig.... Unless you like the additional analog "flavor" of the D/A/D conversion.

Quote:

Getting back to mixing in the box, I'd like to revisit the comment about summing being just math, and then Bob O's reply. If indeed summing is simple addition, then why would this be implemented differently in different DAW's? Addition is addition.

Addition is addition, but it is usually accompanied by multiplication, unless all the DAW's faders are held at exactly 0 dB and pans are full left and right (unless you're Yamaha, but that's another topic). In other words, you'll find that most DAWs manage the addition pretty well, if you run, for example, a 4 into 2 folddown through a DAW with all faders at 0 and if the signal doesn't overload, you'll get a completely transparent summing. You won't even have to add dither!

So, back to multiplication, that does separate the men from the boys, though most would argue that equalization and compression further separates them. The multiplcation should be done at high precision, the coefficients for the fader values well chosen, etc. I'm sure that's done very well within Pro Tools TDM or HD, by the way.

Quote:

And I do hear differences between DAW's. Often very very subtle, but there nevertheless. Even playing back two track stereo mixes between different software can produce slightly differnet tone or levels.

Yes, often subtly different. Makes you bang your head against the wall, as it is DIGITAL. Separating the variables and finding out WHY they sound different is a much more difficult investigation.

Even the L2 Native sounds subtly different from the Hardware or TDM  versions. Different calculations, naturally, so I'm slowly dissolving into obscurity......  

BK

[Albert]

Bob said:

"Now I think this is the reaction of someone who has not experienced the most transparent of SRC conversions. It's been a good 5-10 years since D/A/D conversion sounds better, transparent, more open, and purer than a given SRC. Ask 10 mastering engineers. Ask Bob Ludwig.... Unless you like the additional analog "flavor" of the D/A/D conversion."

Well, then it is obvious I haven't experienced good SRC conversions. Most of what I've heard sounds absolutely terrible. That's using standard settings in DP, Peak, PT, that kind of software. In all fairness, it has been a long time since I've heard BarbaBatch, so it's quite likely more recent versions are better. Perhaps I should try it again.

[jfrigo]

Albert wrote on Tue, 28 December 2004 15:17

Bob said: "Now I think this is the reaction of someone who has not experienced the most transparent of SRC conversions. It's been a good 5-10 years since D/A/D conversion sounds better, transparent, more open, and purer than a given SRC. Ask 10 mastering engineers. Ask Bob Ludwig.... Unless you like the additional analog "flavor" of the D/A/D conversion." Well, then it is obvious I haven't experienced good SRC conversions. Most of what I've heard sounds absolutely terrible. That's using standard settings in DP, Peak, PT, that kind of software. In all fairness, it has been a long time since I've heard BarbaBatch, so it's quite likely more recent versions are better. Perhaps I should try it again.

I must concur with Bob. SRC is no longer a dreaded problem. Sonic HD in the software world, or Lavry, Weiss, and some others in the hardware world, and the current  t.i. 4192 chip (and 4193) being used in some new more affordable designs are all very capable SRC options and are more transparent than a trip through D/A/D. The software you mention is not state of the art for SRC.

OK, only 8 more mastering engineers to ask...

[Ralf Kleemann]

Quote:

You should see what kinds of non-musical garbage above 20 kHz my Cedar Retouch color graph display shows in a number of "96 kHz" recordings. Remember, all the buzz and harmonics of tics and pops and electrical noises end up there along with whatever harmonics of the music might be up there. (...) BK

I'm going back a few postings here, but I just wanted to add that there are enough bad things happening in the audible band - I've heard more than one recording with incredible peaks in the 15-16 kHz region that told me a lot about the listening capabilities of the people involved "Afro Celt Sound System 2: Release" is a good example - possibly an exciter turned up to "11" on the title track, and no one noticed it.

I'm aware that this does not touch the original topic, it's just meant to be a footnote...

Ralf

[danlavry]

jfrigo wrote on Wed, 29 December 2004 07:07

Albert wrote on Tue, 28 December 2004 15:17 Bob said: "Now I think this is the reaction of someone who has not experienced the most transparent of SRC conversions. It's been a good 5-10 years since D/A/D conversion sounds better, transparent, more open, and purer than a given SRC. Ask 10 mastering engineers. Ask Bob Ludwig.... Unless you like the additional analog "flavor" of the D/A/D conversion." Well, then it is obvious I haven't experienced good SRC conversions. Most of what I've heard sounds absolutely terrible. That's using standard settings in DP, Peak, PT, that kind of software. In all fairness, it has been a long time since I've heard BarbaBatch, so it's quite likely more recent versions are better. Perhaps I should try it again. I must concur with Bob. SRC is no longer a dreaded problem. Sonic HD in the software world, or Lavry, Weiss, and some others in the hardware world, and the current  t.i. 4192 chip (and 4193) being used in some new more affordable designs are all very capable SRC options and are more transparent than a trip through D/A/D. The software you mention is not state of the art for SRC. OK, only 8 more mastering engineers to ask...

One needs to “slow down” here. Yes, SRC’s have come a long way. The older AD1890 yielded 106dB THD+N at 1KHz (full scale) and 100dB at 10KHz. The AD1895 specs are much better, THD of 115dB and DR of 125dB. The AD1896 yields 117dB THD and 132dB dynamic range worst case!.  

Crystal (Ciruss) and TI have some great specs as well. Clearly the initial inclination is to go for the better numbers. But what do the numbers mean? The very old AD1890 already had respectable numbers when compared to most IC AD’s (even by today’s standards).
Should there be any difference between say an AD1895 and AD1896?. The story regarding the differences (AD1896 is better) is not so much because of better THD and dynamic range specs.

Say you have a great AD, with a great un weighted dynamic range of 125dB. Say you pass the data through a SRC. What is the combined outcome?
125dB AD with 130B SRC yield 123.8dB dynamics
125dB AD with 135B SRC yield 124.6dB dynamics
125dB AD with 140B SRC yield 123.86dB dynamics
125dB AD with 144B SRC yield 123.946dB dynamics

In other words, there is hardly a difference or improvement.

Let’s repeat it with say 100dB dynamic range digital audio signal (closer to real world 16 bit release)
100dB signal with 130B SRC yield 99.996dB dynamics
100dB signal with 135B SRC yield 99.999dB dynamics
100dB signal with 140B SRC yield 100dB dynamics
100dB signal with 144B SRC yield 100dB dynamics

I could repeat the above exercise for THD+N (total harmonic distortions and noise). With similar OUTCOME NUMBERS for THD+N, one may wonder if and why the better SRC NUMBERS tell the story. I’ll say that much: they don’t tell the whole story.

One has to understand the mechanism of what takes place with the SRC and the clocks. Also one has to understand the measurement itself (what does it measure, and what does it not measure).

But even before we get deep into it, the first question is: How much jitter was there during the measurement? If you wish to use an SRC to help solve jitter, what is the use in specifying it while it is operating with almost no jitter? What will happen when you introduce some jitter?

In my view it is very shallow to see the SRC as a jitter removing magic block. Yes, the devices are getting better, and some are good. But to assume that their operation will not degrade when the input and/or output (sampling ratio) is jittery or varying with time is sloppy at best. Sample rate conversion does alter the data. With more jitter, SRC alters it differently, and some of the tools used to measure do not show the differences well.
 
Regards
Dan Lavry
http://www.lavryengineering.com

“In a time of deceit, telling the truth is a revolutionary act.”

[Timeline]

Curious,

What meathod would one use to 'introduce' controlled jitter for a measurment?
What might the jitter source be?

What might the source be which cause jitter in actual use?

[bobkatz]

Timeline wrote on Sat, 08 January 2005 23:34

Curious, What meathod would one use to 'introduce' controlled jitter for a measurment? What might the jitter source be? What might the source be which cause jitter in actual use?

You use a "jitter generator". I have one of those, actually. But I haven't used it in a long time. In the old days I was trying to get the cleanest signal into a DAC's receiver chip. But that was in the days of the Crystal 8412 receivers and so on. Things have come a long way, dual PLLs implemented well have been done by a select few manufacturers. I no longer bother to measure the interface jitter on the digital side, I just measure the effects on the analog side.

Jitter generators are probably at this point only important to OEMs and chip manufacturers. To test jitter attenuation.

BK

[danlavry]

bobkatz wrote on Sun, 20 February 2005 21:45

Timeline wrote on Sat, 08 January 2005 23:34 Curious, What meathod would one use to 'introduce' controlled jitter for a measurment? What might the jitter source be? What might the source be which cause jitter in actual use? ....I just measure the effects on the analog side. BK

I heard about such practices, the "poor man" way to measure jitter by looking at an FFT. It never made sense to me, and still does not. That 1KHz sideband measurement around 10KHz tone does not reveal much of the jitter picture. Generating those 1KHz sideband requires a well calibrated 1KHz modulated jitter source, and the FFT test is only about the rejection of 1KHz energy. It does not show all the random causes of jitter, such as an oscillator with a lot of random jitter energy (say below 1KHz), or other jitter sources (such as power line coupling or coupling of the MSB from the data to the clock, which can induce a lot of jitter below 1KHz.

So far, that FFT 1KHz sideband test seems to me to have been invented by some one's marketing department, but I am certainly open and ready to be corrected.

It is not only the concept that I question. The measurement itself may yield some approximate results while inducing high level of jitter, but as you approach tiny jitter, the measurement loses accuracy. One can not assume that injecting say 1nsec 1kHZ modulation will yield information about what will happen when you inject say 50psec jitter. The random jitter and the 1KHz "forced" jitter do not add linearly. Not to mention that the test does not answer the jitter rejection at across the audio band and way beyond (jitter at any frequency is bad news).

I do not view the job of inserting say a 50psec 1KHz jitter, with say 10psec random component, a poor man tool. I can design and build such a source, and verifying (measuring) it will take some real jitter meaqsuring equipment - we are talking upwards of $50000 at those low levels.

If one wishes to learn about the PLL filter characteristics, I would sugest that there are simpler and very solid methods available, and they yield a plot across a wide frequency band, not just 1KHz. Of measuring the filter does not tell the whole story either.

The costs of state of the art jitter measurement equipment are very high, but I do not view it as a justification for cutting corners...

I hope I am wrong. I also want to find a cheaper way to measure jitter.

Regards
Dan Lavry
www.lavryenineering.com

     

[bobkatz]

danlavry wrote on Tue, 22 February 2005 11:29

bobkatz wrote on Sun, 20 February 2005 21:45 Timeline wrote on Sat, 08 January 2005 23:34 Curious, What meathod would one use to 'introduce' controlled jitter for a measurment? What might the jitter source be? What might the source be which cause jitter in actual use? ....I just measure the effects on the analog side. BK I heard about such practices, the "poor man" way to measure jitter by looking at an FFT. It never made sense to me, and still does not.

Well, ultimately it is the effect on the analog signal that counts. Jitter in the interface is a bit like the "if a tree falls in the forest does anyone hear it" argument, because until it goes through the PLL and is attenuated, and then is reclocked, drives the converter section and then the "modulated" affect can be seen or heard... what's the point?

The late Julian Dunn invented the J-Test signal which has been very successful. And a simple 1/4 sample rate sine wave test signal does reveal differences between various converters quite well. In addition, TC Electronic has written an excellent white paper on a simple 12 kHz sine wave test:

Please see:

http://www.tcelectronic.com/default.asp?id=1573  and click on "Clock and Synchronization inthe System 6000"

Quote:

It is not only the concept that I question. The measurement itself may yield some approximate results while inducing high level of jitter, but as you approach tiny jitter, the measurement loses accuracy.

I agree, when you're interested in looking at the clock directly. But accuracy compared to what? Isn't it ultimately the effect on the analog stage that you are looking for? Yes, designers like you absolutely need to be able to measure the clock jitter directly and the attenuation of the PLL.  If they have the money to purchase a clock jitter analyser, as you say, not a poor man's tool) but ultimately it is the effect on the analog signal that counts. And the job of the testing engineer to see the effect on the analog signal. How little jitter do you want to have in a clock signal? The answer: As little jitter as is necessary so that it has no audible (or measurable) affect on the analog signal.

[danlavry]

[quote title=bobkatz wrote on Wed, 23 February 2005 13:23][quote title=danlavry wrote on Tue, 22 February 2005 11:29][quote title=bobkatz wrote on Sun, 20 February 2005 21:45]

Timeline wrote on Sat, 08 January 2005 23:34

Curious, What meathod would one use to 'introduce' controlled jitter for a measurment? What might the jitter source be? The late Julian Dunn invented the J-Test signal which has been very successful. And a simple 1/4 sample rate sine wave test signal does reveal differences between various converters quite well. In addition, TC Electronic has written an excellent white paper on a simple 12 kHz sine wave test: Please see: http://www.tcelectronic.com/default.asp?id=1573  and click on "Clock and Synchronization inthe System 6000" Quote: It is not only the concept that I question. The measurement itself may yield some approximate results while inducing high level of jitter, but as you approach tiny jitter, the measurement loses accuracy. If they have the money to purchase a clock jitter analyser, as you say, not a poor man's tool) but ultimately it is the effect on the analog signal that counts. And the job of the testing engineer to see the effect on the analog signal. How little jitter do you want to have in a clock signal? The answer: As little jitter as is necessary so that it has no audible (or measurable) affect on the analog signal.

I am well familiar with Dunns paper and did read the TC paper as well. I am even more fimiliar with the math of what it takes to make those 1KHz FFT spikes. It takes a 1KHz sine wave to modulate the phase of a 10KHz (or 12KHZ for the TC paper), and the amplitude has to be pretty high. It sorts of work to indicate the rejection at 1KHz as long as the FFT is long, the FFT window is good and the 1KHz jitter is much higher then the real system jitter. I could not resolve to any degree of accuracy accurately a simulation of 100psec 1KHz sine wave jitter on an 8K FFT with BH4 window. That was a simulation, NO real noise. NO random noise!

So that "tool" is great for instructive purposes, but not for measuring real jitter. Unfortunately, some folks are "marketing" their gear as low jitter based on such measurements.

Again, that measurement becomes very inaccurate at low jitter levels, and does not tell you information about anything outside of the loop rejection at 1KHz to relatively high jitter, which by itself is of little value, because it ignores the rest of the picture.

Also, it is not difficult to build a jitter generator for inducing a 10KHz digital tone with 3.5nsec jitter, because the presence of say 100psec random wide band jitter it relatively negligible.

But try to build a 10KHz generator with say 50psec of 1KHz tone. This time, you can not afford 100psec random jitter, and even 25psec random jitter is undesirable...

So the jitter generator is a problem. Injecting low level jitter is a problem too. The FFT does not read low level jitter accurately. The information is restricted to 1KHz only... Other then that it is a great jitter test!

The way to measure low level jitter is to have an extremely wide band measurement gear! I read today that Agilent makes a probe with 220fF (0.22pF) capacitance!  
 
Of course I agree that the end result is what counts, and that less jitter is better. I am sad to see people in audio make such a big deal out of that almost worthless test. I think it is being used for just more marketing hype. You called the test "successful", I disagree. We seem to agree that real jitter measurement tools are very expansive! Some of it way upward of $50000! I just do not see the expanse as justification or a "permit" to come up with an inadequate and inconclusive "test".  

Regards
Dan Lavry
www.lavryengineering.com

[bobkatz]

Dear Dan:

I cannot speak mathematically to the test's resolution , but I can say that the J-test has proved extremely revealing for me as far as the jitter is concerned in a D/A converter. There are DACs which measure terribly, and there are a handful of DACs which show no artifacts at all, when measuring with a 64k point FFT, averaging over about 4 seconds, with 64 Bit float accuracy, on a 44.1 kHz sampled signal.

Have you read any of John Atkinson's measurements using the J-Test signal and a special analyser that really just identifies the spectral products by their distance and how they surround the 11.05 kHz fundamental?

I can send you various measurements I have made of D/A and A/D converters. The cheap ones measure worse, for sure! I can find a small quantity (less than a handful) of DACs whose jitter artifacts are equal to or below their noise floor! What does that tell you?

I personally found the signal to be very informative and correlate with the sonics of various converters that are/are not susceptible to jitter. I have measured one DAC which uses the TI ASRC chip but in the first design was driven from a defective crystal. This same DAC shows EXTREME improvement (no artifacts in the FFT) after the crystal was fixed.

I wish I could speak to magnitudes of jitter, but as I said, Atkinson's measurements show quantities that are inferred from this indirect method of measurement.

Since a ladder converter and a delta-sigma converter perform quite differently with respect to a jittery clock (see Bob Adam's papers) how do you go about evaluating the sonic effects of the jitter without taking into account the effects on the analog side? In other words, a true ladder DAC can be fed much higher levels of jitter and exhibit far fewer artifacts on the analog side than a delta-sigma DAC.

Quote:

But try to build a 10KHz generator with say 50psec of 1KHz tone. This time, you can not afford 100psec random jitter, and even 25psec random jitter is undesirable...

That's the hard part, of course, building an analog generator with little jitter itself for the A/D converter under test! The Audio Precision that I borrow occasionally (old model, analog only, Portable one) does not have low enough THD! But the generator in the Audio Toolbox seems to have very low artifacts...